Fix seed sensitivity of test_fastica_eigh_low_rank_warning (scikit-learn#29612)

Author: ogrisel

doc/whats_new/v1.6.rst

@@ -165,6 +165,14 @@ Changelog
   and automatic retries in case of HTTP errors. :pr:`29354` by :user:`Olivier
   Grisel <ogrisel>`.
 
+:mod:`sklearn.decomposition`
+............................
+
+- |Fix| Increase rank defficiency threshold in the whitening step of
+  :class:`decomposition.FastICA` with `whiten_solver="eigh"` to improve the
+  platform-agnosticity of the estimator. :pr:`29612` by :user:`Olivier Grisel
+  <ogrisel>`.
+
 :mod:`sklearn.discriminant_analysis`
 ....................................
 

sklearn/decomposition/_fastica.py

@@ -605,7 +605,7 @@ def g(x, fun_args):
                 # Faster when num_samples >> n_features
                 d, u = linalg.eigh(XT.dot(X))
                 sort_indices = np.argsort(d)[::-1]
-                eps = np.finfo(d.dtype).eps
+                eps = np.finfo(d.dtype).eps * 10
                 degenerate_idx = d < eps
                 if np.any(degenerate_idx):
                     warnings.warn(

sklearn/decomposition/tests/test_fastica.py

@@ -13,7 +13,7 @@
 from sklearn.decomposition import PCA, FastICA, fastica
 from sklearn.decomposition._fastica import _gs_decorrelation
 from sklearn.exceptions import ConvergenceWarning
-from sklearn.utils._testing import assert_allclose
+from sklearn.utils._testing import assert_allclose, ignore_warnings
 
 
 def center_and_norm(x, axis=-1):
@@ -448,5 +448,10 @@ def test_fastica_eigh_low_rank_warning(global_random_seed):
     X = A @ A.T
     ica = FastICA(random_state=0, whiten="unit-variance", whiten_solver="eigh")
     msg = "There are some small singular values"
+
     with pytest.warns(UserWarning, match=msg):
-        ica.fit(X)
+        with ignore_warnings(category=ConvergenceWarning):
+            # The FastICA solver may not converge for some data with specific
+            # random seeds but this happens after the whiten step so this is
+            # not want we want to test here.
+            ica.fit(X)